v0.341.0: the night update takes a security_fix ladder step (D39 = B, 09 §3 decision 255, R-1239); every other leg guard unchanged; the leg names the CVE ids; MinAgent 0.158.1
gates / gates (push) Successful in 1m29s

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012qRErfCoiTkvDK9N5XHbzb
This commit is contained in:
2026-10-23 13:09:52 +02:00
parent 32e61d5d1d
commit 63dcccfeec
7 changed files with 191 additions and 30 deletions
+22
View File
@@ -1,3 +1,25 @@
## v0.341.0 (2026-10-23) — the night update takes a security-fix step (D39 = B, `09` §3 decision 255, R-1239)
**MinAgent: 0.158.1** (unchanged). No hub dependency (the hub report's `update_leg.steps[].security` is a new optional
field; the hub stores the report as it comes). Task: the D39 = B day brief (attended).
- **`legCandidate`** takes a ladder step whose verdict is `proven` OR `security_fix` (`stacks.LadderVerdictSecurityFix`) —
an urgent upstream fix proven on the box (scratch 9202, the guarded Update), its bench run optional, its public notice in
the entry's `security` block (CVE/GHSA ids, the advisory, the first fixed version; `stacks.LadderSecurity`). Every other
guard is unchanged and runs in the same order: held, stopped by the household, current/ahead/unorderable, unpinned, no
ladder, older than the ladder, `needs_person`, the failed step (R-680), `files_may_change` without a whole copy, W+5h.
- The leg logs `[update-leg] <app>: security fix <ids> taken (<from> → <to>)` when it presses such a step (beside the
usual `step pressed`); the summary step carries `security: [ids]` (hub report) and the summary line ends
`[security fix <app>: <ids> <outcome>]`. A skipped security step logs its reason like any skip.
- `LadderEntry.Security` is `omitempty`, so `LadderPrint` (R-680's fingerprint) of every existing entry is unchanged.
- **Unchanged:** a person's press (it never read the verdict — on 0.340.0 a hand press took the step; proven on 9202),
the guarded update, the badge.
- Tests (`internal/stacks/unattended_security_test.go`, the real leg over the real guarded update): taken and named;
`needs_person` skipped; held skipped; stopped by the household skipped; an unknown verdict still `not_proven`; `proven`
unchanged with no security names. Red-proofs: the OLD line turns the first two red; „a security_fix returns before the
guards" turns needs_person/held/stopped red; „any verdict but unrecorded is taken" turns the unknown-verdict test red;
„log every step as a security fix" turns the proven test red.
## v0.340.0 (2026-10-23) — the night copy in a capped container that never outlives the controller (R-1214 + the stray-copy fix), Immich's model cache out of backups (R-49, D21 = B), no import over a stopped app (R-1187 c), the catch-up's checks (R-1189 d, C5) ## v0.340.0 (2026-10-23) — the night copy in a capped container that never outlives the controller (R-1214 + the stray-copy fix), Immich's model cache out of backups (R-49, D21 = B), no import over a stopped app (R-1187 c), the catch-up's checks (R-1189 d, C5)
**MinAgent: 0.158.1** (unchanged). No hub dependency. Task: the 2026-10-23 day brief (release run, operator go). Soaked **MinAgent: 0.158.1** (unchanged). No hub dependency. Task: the 2026-10-23 day brief (release run, operator go). Soaked
+6
View File
@@ -1,5 +1,11 @@
# CONTEXT.md — Project Memory # CONTEXT.md — Project Memory
> **2026-10-23 (day, later) — v0.341.0: the night leg takes a `security_fix` ladder step (D39 = B, `09` §3 decision 255,
> R-1239; operator ruling by starting the run).** `legCandidate`: verdict `proven` OR `security_fix`, every other guard
> unchanged and in order; log `[update-leg] <app>: security fix <ids> taken (<from> → <to>)`; `LegStep.Security`. The
> catalog side (format, gates, writer, `box_walk.py --security`) is catalog `3898593`. On ≤ 0.340.0 the leg skips such a
> step (`not_proven`) and a hand press takes it — both proven on 9202 (felhom.eu `audits/security-lane-2026-10-23/`).
> **2026-10-22/23 night — 0.340.0 candidate on `main`, unreleased.** The stray-copy fix (`66a69d4`, `912e592`) and two > **2026-10-22/23 night — 0.340.0 candidate on `main`, unreleased.** The stray-copy fix (`66a69d4`, `912e592`) and two
> review rounds (`943c517` C5/C3/C6/C8; `9d9e404` N2–N5, N9; `1babfe9` two races): the capped copy's container has a > review rounds (`943c517` C5/C3/C6/C8; `9d9e404` N2–N5, N9; `1babfe9` two races): the capped copy's container has a
> fixed name + label, is removed by name on any error, leftovers are removed at start before the scheduler/web server > fixed name + label, is removed by name on any error, leftovers are removed at start before the scheduler/web server
+10 -25
View File
@@ -1,28 +1,13 @@
# REPORT — day 2026-10-23: controller 0.340.0 released (the operator's „go") # REPORT — v0.341.0: the night update takes a security-fix step (2026-10-23, D39 = B, `09` §3 decision 255)
**What:** the day brief's Part 1. Design read: `07-backup-architecture.md` §6.3, §8; `09-update-architecture.md` §3 **What:** `legCandidate` takes `proven` OR `security_fix`; every other guard unchanged and in the same order; the leg
decision 250 (a release needs no golden; the floor carries it, MinAgent declared). names a taken security step in its log, its summary step and its summary line. Design read: `09-update-architecture.md`
§3 decisions 13–15, 52, 255.
## Release **Tests** (`internal/stacks/unattended_security_test.go`, six): red-proved by four mutations — the OLD verdict line
(taken + needs_person red), a security_fix returning before the guards (needs_person, held, stopped red), any verdict but
unrecorded taken (unknown-verdict red), every step logged as a security fix (proven red). Held and stopped run BEFORE the
verdict line, so the old line cannot turn them red; that is why their red-proof is a different mutation.
`go build`, `go vet`, `go test ./...` green; `controller_gates.py --fast` OK.
- Main at `caaa86e` (the CHANGELOG header `## v0.340.0 … MinAgent: 0.158.1`; the four night sections now under it). **Live:** see the felhom.eu evidence `documentation/audits/security-lane-2026-10-23/` (P2 old controller, P3 the rc).
CI 2245 success. The operator answered „release main" for the text-only commits after the night-tested ones.
- Image `gitea.dooplex.hu/admin/felhom-controller:0.340.0`, digest
`sha256:a8d1c6458da4ea9d667f928498656a291771e0b3756b20f2fcc8c1cf9d516621`.
- 9202 smoke first (healthy, 0 errors, `/login` 200, start-up „no leftover").
- Per-customer floors for Tester 1, demo-hp, demo-felhom (303 `floor_set` each) → all three on 0.340.0 through the
normal update path: Tester 1 10 containers, demo-hp 24, N100 9; 0 unhealthy, 0 errors, 0 copy containers, each with the
start-up „no leftover" line. `https_only_check.py` rc 0, 28 OK, the negative control convicted.
- „Fájlok ↗" on Tester 1: `/__files/open` → 302 to `files.enkicsifelhom.hu/_felhom/sso`. Tester 1's manual second-drive
run: 3 apps, 0 copy containers left. demo-hp and demo-felhom: checked only without a session (302 /login) — no
dashboard password is held for them; their first night copy is tonight.
- Global floor 0.340.0, MinAgent 0.158.1: 303 `floor_set` at 06:41:04 UTC; the per-customer floors cleared; read back
from the hub DB with its WAL.
## What it carries
The stray-copy fix and the night's review fixes (R-1214, R-49 D21 = B, R-1187 c, R-1189 d, C3–C8, N2–N5, N9, S1, S2) —
see `CHANGELOG.md` v0.340.0. R-49 closed; R-1214 stays open for one dated check (2026-10-24: Tester 1's first night copy
in the capped container, memory „full" pressure under 20 %); R-1187 and R-1189 keep their unfixed parts.
Evidence: felhom.eu `documentation/audits/day-2026-10-23/01`–`06`. Full day: felhom.eu `REPORT-day-2026-10-23.md`.
+8
View File
@@ -873,6 +873,14 @@ sends no mail; the app page says „Automatikus frissítés %s-kor — sikeres."
One summary line per night (`[update-leg] update leg (after-offsite): done=… undone=… held=… failed=… One summary line per night (`[update-leg] update leg (after-offsite): done=… undone=… held=… failed=…
skipped=… [skipped: app=reason, …]`) in the log and in the hub report's `update_leg`. After `done`/`undone` the skipped=… [skipped: app=reason, …]`) in the log and in the hub report's `update_leg`. After `done`/`undone` the
app's steps-left and badge are fresh at once (R-678). `stacks.update_window` is removed (it was never read). app's steps-left and badge are fresh at once (R-678). `stacks.update_window` is removed (it was never read).
**A security-fix step (v0.341.0, `09` §3 decision 255, D39 = B).** The leg also takes a ladder entry with
`"verdict": "security_fix"` — an urgent upstream fix proven on the box (scratch 9202, the guarded Update) whose bench run
is optional, with its public notice (`security`: CVE/GHSA ids, the advisory, the first fixed version). Every other rule
above holds for it unchanged and in the same order (held, stopped by the household, `needs_person`, the failed step,
`files_may_change`, W+5h). The leg logs `[update-leg] <app>: security fix <ids> taken (<from> → <to>)`, the hub
report's step carries `security: [ids]`, and the summary line ends `[security fix <app>: <ids> <outcome>]`. A
controller before 0.341.0 skips such a step at night (`not_proven`); a person's press takes it on every controller
with ladders (the press never read the verdict).
**Leftovers (v0.272.0).** The backup page says in plain words when a whole-box backup was skipped because it **Leftovers (v0.272.0).** The backup page says in plain words when a whole-box backup was skipped because it
cannot fit (the agent's numbers; R-685). A restore that lifts an update hold removes that hold's undo copies cannot fit (the agent's numbers; R-685). A restore that lifts an update hold removes that hold's undo copies
+16
View File
@@ -52,6 +52,22 @@ type LadderEntry struct {
// PostgreSQL major (pgconvert.go). omitempty: an entry without it prints exactly as before, so R-680's // PostgreSQL major (pgconvert.go). omitempty: an entry without it prints exactly as before, so R-680's
// failed-step records (tied to LadderPrint) are not reset by this field's arrival. // failed-step records (tied to LadderPrint) are not reset by this field's arrival.
EngineConversion *EngineConversion `yaml:"engine_conversion" json:"engine_conversion,omitempty"` EngineConversion *EngineConversion `yaml:"engine_conversion" json:"engine_conversion,omitempty"`
// Security (v0.341.0, `09` §3 decision 255, D39 = B) is a `security_fix` step's public notice: the CVE/GHSA ids,
// the advisory's address and the first fixed upstream version. The night leg names the ids when it takes the step.
// omitempty, like EngineConversion: an entry without it prints (LadderPrint) exactly as before.
Security *LadderSecurity `yaml:"security" json:"security,omitempty"`
}
// LadderVerdictSecurityFix is the second verdict the night leg takes (v0.341.0, decision 255): an urgent upstream fix
// proven on the box (scratch 9202, the guarded Update) whose bench run is optional. A person's press has never read the
// verdict, so on every controller with ladders it was already pressable by hand; before v0.341.0 the leg skips it.
const LadderVerdictSecurityFix = "security_fix"
// LadderSecurity is the `security` object of a security_fix ladder entry (the catalog's `scripts/ladder.py` checks it).
type LadderSecurity struct {
IDs []string `yaml:"ids" json:"ids"`
Advisory string `yaml:"advisory" json:"advisory"`
FixedIn string `yaml:"fixed_in" json:"fixed_in"`
} }
// LadderMarks is the `marks` object of a ladder entry. NeedsPerson is JSON null (nil) or the tester's // LadderMarks is the `marks` object of a ladder entry. NeedsPerson is JSON null (nil) or the tester's
+21 -5
View File
@@ -19,7 +19,8 @@ import (
// It presses the SAME public guarded Update a person presses (StartGuardedUpdate) — no second update // It presses the SAME public guarded Update a person presses (StartGuardedUpdate) — no second update
// path — one app at a time, ONE tested step per app per night, and never: // path — one app at a time, ONE tested step per app per night, and never:
// - an app whose installed version has no ladder entry (older than the ladder, or no test record); // - an app whose installed version has no ladder entry (older than the ladder, or no test record);
// - a step whose verdict is not `proven`, or that carries the `needs_person` mark; // - a step whose verdict is not `proven` or `security_fix` (v0.341.0, decision 255 — an urgent fix proven on the
// box, bench optional; every other guard below holds for it unchanged), or that carries the `needs_person` mark;
// - a `files_may_change` step unless a fresh WHOLE copy of the app exists on the box (decision 25/26); // - a `files_may_change` step unless a fresh WHOLE copy of the app exists on the box (decision 25/26);
// - the step that was undone or held last time, while the catalog's ladder is unchanged (R-680); // - the step that was undone or held last time, while the catalog's ladder is unchanged (R-680);
// - a held app, an app that is current, ahead, or unorderable. // - a held app, an app that is current, ahead, or unorderable.
@@ -92,6 +93,8 @@ type LegStep struct {
From map[string]string `json:"from,omitempty"` From map[string]string `json:"from,omitempty"`
To map[string]string `json:"to,omitempty"` To map[string]string `json:"to,omitempty"`
Seconds float64 `json:"seconds,omitempty"` Seconds float64 `json:"seconds,omitempty"`
// Security names the public notice of a security_fix step the leg took (v0.341.0, decision 255): its CVE/GHSA ids.
Security []string `json:"security,omitempty"`
} }
// UpdateLegSummary is one night's leg — the operator's summary line and the hub report's `update_leg`. // UpdateLegSummary is one night's leg — the operator's summary line and the hub report's `update_leg`.
@@ -123,9 +126,15 @@ func (s *UpdateLegSummary) Line() string {
if s.Stopped != "" { if s.Stopped != "" {
stop = " stopped=" + s.Stopped stop = " stopped=" + s.Stopped
} }
return fmt.Sprintf("update leg (%s): done=%d undone=%d held=%d failed=%d skipped=%d%s in %s [skipped: %s]", sec := ""
for _, st := range s.Steps {
if len(st.Security) > 0 && st.Outcome != LegOutcomeSkipped {
sec += fmt.Sprintf(" [security fix %s: %s %s]", st.App, strings.Join(st.Security, ","), st.Outcome)
}
}
return fmt.Sprintf("update leg (%s): done=%d undone=%d held=%d failed=%d skipped=%d%s in %s [skipped: %s]%s",
s.Trigger, s.Done, s.Undone, s.Held, s.Failed, s.Skipped, stop, s.Trigger, s.Done, s.Undone, s.Held, s.Failed, s.Skipped, stop,
s.EndedAt.Sub(s.StartedAt).Round(time.Second), strings.Join(skips, ", ")) s.EndedAt.Sub(s.StartedAt).Round(time.Second), strings.Join(skips, ", "), sec)
} }
func (s *UpdateLegSummary) add(st LegStep) { func (s *UpdateLegSummary) add(st LegStep) {
@@ -334,10 +343,17 @@ func (m *Manager) runUpdateLeg(ctx context.Context, trigger string, manual bool)
continue continue
} }
m.logger.Printf("[INFO] [update-leg] %s: step pressed %s → %s", name, summarisePin(entry.From), summarisePin(entry.To)) m.logger.Printf("[INFO] [update-leg] %s: step pressed %s → %s", name, summarisePin(entry.From), summarisePin(entry.To))
var secIDs []string
if entry.Verdict == LadderVerdictSecurityFix {
if entry.Security != nil {
secIDs = entry.Security.IDs
}
m.logger.Printf("[INFO] [update-leg] %s: security fix %s taken (%s → %s)", name, strings.Join(secIDs, ","), summarisePin(entry.From), summarisePin(entry.To))
}
m.setLegFlags(true, true) m.setLegFlags(true, true)
outcome := m.legWait(ctx, o, name) outcome := m.legWait(ctx, o, name)
m.setLegFlags(true, false) m.setLegFlags(true, false)
step := LegStep{App: name, Outcome: outcome, From: entry.From, To: entry.To, Seconds: o.Now().Sub(t0).Round(100 * time.Millisecond).Seconds()} step := LegStep{App: name, Outcome: outcome, From: entry.From, To: entry.To, Seconds: o.Now().Sub(t0).Round(100 * time.Millisecond).Seconds(), Security: secIDs}
sum.add(step) sum.add(step)
m.recordAutoUpdate(name, &AutoUpdateRecord{At: t0.UTC().Format(time.RFC3339), Outcome: outcome, From: entry.From, To: entry.To}) m.recordAutoUpdate(name, &AutoUpdateRecord{At: t0.UTC().Format(time.RFC3339), Outcome: outcome, From: entry.From, To: entry.To})
m.logger.Printf("[INFO] [update-leg] %s: step ended %s after %.1f s", name, outcome, step.Seconds) m.logger.Printf("[INFO] [update-leg] %s: step ended %s after %.1f s", name, outcome, step.Seconds)
@@ -438,7 +454,7 @@ func (m *Manager) legCandidate(ctx context.Context, name string, o *UpdateLegOpt
return LadderEntry{}, LegSkipOlderThanLadder return LadderEntry{}, LegSkipOlderThanLadder
} }
e := ladder[idx] e := ladder[idx]
if e.Verdict != "proven" { if e.Verdict != "proven" && e.Verdict != LadderVerdictSecurityFix {
return e, LegSkipNotProven return e, LegSkipNotProven
} }
if e.Marks.NeedsPerson != nil && strings.TrimSpace(*e.Marks.NeedsPerson) != "" { if e.Marks.NeedsPerson != nil && strings.TrimSpace(*e.Marks.NeedsPerson) != "" {
@@ -0,0 +1,108 @@
package stacks
import (
"strings"
"testing"
)
// v0.341.0 — D39 = B (`09` §3 decision 255): the night leg takes a `security_fix` step like a proven one, and every
// other guard holds for it unchanged and in the same order. Same real leg over the real guarded update as
// unattended_test.go (legManager: nextcloud pinned at A, the catalog at C, ladder A→B→C).
//
// COMPANION RED-PROOF (REPORT.md): against the OLD verdict line (`e.Verdict != "proven"`) the first two tests fail
// (the step is skipped as not_proven). The held and stopped-by-household guards run BEFORE the verdict line, so the old
// line cannot turn them red; their red-proof is the mutation "a security_fix step returns before the guards" (both
// fail). The unknown-verdict test's red-proof is the mutation "any verdict but unrecorded is taken" (it fails).
const secMarks = `{"files_may_change": false, "needs_person": null, "memory_tight": false}`
// securityLine is a security_fix ladder entry A→B with its public notice, optionally with other marks.
func securityLine(from, to, marks string) string {
return ` - {"from": {"web": "` + from + `"}, "to": {"web": "` + to + `"}, "digest": {"web": "sha256:` +
strings.Repeat("a", 64) + `"}, "verdict": "security_fix", "marks": ` + marks +
`, "security": {"ids": ["CVE-2026-96652", "CVE-2026-96654"], "advisory": "https://www.cisa.gov/kev", "fixed_in": "1.43.3.10861"}, "seed": "none"}` + "\n"
}
func TestD39_SecurityFixStepIsTakenAndNamed(t *testing.T) {
m, dir, _, ups, logBuf := legManager(t)
writeLadder(t, m, securityLine(ladderA, ladderB, secMarks), ladderLine(ladderB, ladderC))
legOpts(m, nil)
s := mustLeg(t, m)
if len(*ups) != 1 || pinOf(t, dir) != ladderB {
t.Fatalf("the security_fix step was not taken: ups=%v pin=%s summary=%+v", *ups, pinOf(t, dir), s)
}
st := legStepFor(s, "nextcloud")
if st.Outcome != LegOutcomeDone || strings.Join(st.Security, ",") != "CVE-2026-96652,CVE-2026-96654" {
t.Fatalf("the summary step must be done and name the ids, got %+v", st)
}
want := "[update-leg] nextcloud: security fix CVE-2026-96652,CVE-2026-96654 taken (" + summarisePin(map[string]string{"web": ladderA}) + " → " + summarisePin(map[string]string{"web": ladderB}) + ")"
if !strings.Contains(logBuf.String(), want) {
t.Fatalf("the log must carry %q; log:\n%s", want, logBuf.String())
}
if !strings.Contains(s.Line(), "[security fix nextcloud: CVE-2026-96652,CVE-2026-96654 done]") {
t.Fatalf("the operator's summary line must name the security fix: %s", s.Line())
}
}
func TestD39_SecurityFixWithNeedsPersonIsSkipped(t *testing.T) {
m, dir, _, ups, _ := legManager(t)
writeLadder(t, m, securityLine(ladderA, ladderB, `{"files_may_change": false, "needs_person": "the admin must sign in again", "memory_tight": false}`),
ladderLine(ladderB, ladderC))
legOpts(m, nil)
s := mustLeg(t, m)
if len(*ups) != 0 || pinOf(t, dir) != ladderA || legStepFor(s, "nextcloud").Reason != LegSkipNeedsPerson {
t.Fatalf("security_fix + needs_person: ups=%v pin=%s step=%+v, want skipped %q", *ups, pinOf(t, dir), legStepFor(s, "nextcloud"), LegSkipNeedsPerson)
}
}
func TestD39_SecurityFixOnAHeldAppIsSkipped(t *testing.T) {
m, _, g, ups, _ := legManager(t)
g.held, g.holdWhy = true, "HELD"
writeLadder(t, m, securityLine(ladderA, ladderB, secMarks), ladderLine(ladderB, ladderC))
legOpts(m, nil)
s := mustLeg(t, m)
if len(*ups) != 0 || legStepFor(s, "nextcloud").Reason != LegSkipHeld {
t.Fatalf("security_fix on a held app: ups=%v step=%+v, want skipped %q", *ups, legStepFor(s, "nextcloud"), LegSkipHeld)
}
}
func TestD39_SecurityFixOnAnAppTheHouseholdStoppedIsSkipped(t *testing.T) {
m, _, _, ups, _ := legManager(t)
writeLadder(t, m, securityLine(ladderA, ladderB, secMarks), ladderLine(ladderB, ladderC))
legOpts(m, nil)
if err := m.ScanStacks(); err != nil {
t.Fatal(err)
}
if err := m.SetDesiredState("nextcloud", DesiredStateStopped); err != nil {
t.Fatal(err)
}
s := mustLeg(t, m)
if len(*ups) != 0 || legStepFor(s, "nextcloud").Reason != LegSkipStoppedByHousehold {
t.Fatalf("security_fix on a stopped app: ups=%v step=%+v, want skipped %q", *ups, legStepFor(s, "nextcloud"), LegSkipStoppedByHousehold)
}
}
func TestD39_AnUnknownVerdictIsStillNotProven(t *testing.T) {
for _, v := range []string{"security-fix", "failed", "inconclusive", ""} {
m, _, _, ups, _ := legManager(t)
writeLadder(t, m, ladderLineMarks(ladderA, ladderB, v, secMarks), ladderLine(ladderB, ladderC))
legOpts(m, nil)
s := mustLeg(t, m)
if len(*ups) != 0 || legStepFor(s, "nextcloud").Reason != LegSkipNotProven {
t.Fatalf("verdict %q: ups=%v step=%+v, want skipped %q", v, *ups, legStepFor(s, "nextcloud"), LegSkipNotProven)
}
}
}
func TestD39_ProvenIsUnchangedAndCarriesNoSecurityNames(t *testing.T) {
m, dir, _, _, logBuf := legManager(t)
legOpts(m, nil)
s := mustLeg(t, m)
st := legStepFor(s, "nextcloud")
if pinOf(t, dir) != ladderB || st.Outcome != LegOutcomeDone || st.Security != nil {
t.Fatalf("a proven step: pin=%s step=%+v, want B done with no security names", pinOf(t, dir), st)
}
if strings.Contains(logBuf.String(), "security fix") || strings.Contains(s.Line(), "security fix") {
t.Fatalf("a proven step must not be logged as a security fix")
}
}